关于改善地面穿透雷达图像数据失衡的方法的研究
Ligang Cao1, Lei Liu2, Congde Lu1
1Key Laboratory of Earth Exploration and Information Techniques of Ministry of Education, Chengdu University of Technology, Chengdu, 610059, Sichuan, China.
Scientific reports
|January 22, 2025
概括
本研究介绍了一种成本有效的方法,使用生成对抗网络来扩大道路损坏数据集,用于地面透雷达 (GPR) 分析. 该方法提高了对路面损坏分类的深度学习准确性,即使初始数据有限.
科学领域:
- 地质学和铺路工程 铺路工程
- 人工智能和机器学习
背景情况:
- 地面透雷达 (GPR) 对于高速公路路面的评估至关重要.
- 深度学习对GPR数据解释有希望,但由于道路损坏数据集的不平衡,导致分类准确性差.
- 与收集大量现场数据相关的高成本限制了当前方法的可扩展性.
研究的目的:
- 提出一种成本效益高的数据扩展方法,用于使用GPR进行道路损坏分类.
- 提高用于标识路面损坏的深度学习模型的准确性.
- 为了应对道路交通事故分类中数据集不平衡的挑战.
主要方法:
- 开发了一种基于集编码器的生成对抗网络 (GAN) 的创新损害数据扩展方法.
- 对GAN的发电机,区分器和编码器组件进行了具体的改进.
- 评估了不同样本比例的影响,并与传统扩展方法进行了比较.
主要成果:
- 提出的基于GAN的方法成功地从有限的初始数据中生成了稳定和多样化的损伤样本.
- 生成器,区分器和编码器的改进提高了由指标验证的生成样本的质量.
- 该方法在分类网络准确性和频率初始距离 (FID) 得分方面取得了显著的改进.
结论:
- 提出的基于生成对抗网络的数据扩展方法有效地解决了道路损坏分类中不平衡的数据集.
- 这种方法为广泛的现场数据收集提供了低成本的替代方案,同时提高了深度学习模型的性能.
- 增强的GAN架构提供了一个强大的解决方案,用于生成现实的路面损坏样本,用于GPR分析.
更多相关视频
11:37RGB and Spectral Root Imaging for Plant Phenotyping and Physiological Research: Experimental Setup and Imaging Protocols
Published on: August 8, 2017
16.1K
12:03Two-Dimensional Visualization and Quantification of Labile, Inorganic Plant Nutrients and Contaminants in Soil
Published on: September 1, 2020
6.0K
相关概念视频
01:12Taping Over Different Ground Profiles
23
Taping over varying ground profiles requires careful adaptation to achieve accurate measurements. On smooth, level ground with minimal vegetation, the tape can rest directly on the ground. Here, the taping team, typically consisting of a head and a rear tapeman, coordinates their positions with clear communication. The rear tapeman holds the tape at the starting point and guides the head tapeman toward a range pole placed beyond the endpoint, using hand or voice signals to ensure alignment.On...
23
01:17Common Leveling Mistakes and Errors
59
A survey team is tasked with determining the elevation difference between points Point A and Point B, separated by uneven terrain. They use a leveling instrument and a leveling rod.Common MistakesMisreading the Rod: During a backsight reading at Point A, the instrumentman observes the rod partially obscured by tall grass. Instead of reading 1.135 m, they mistakenly record 1.735 m due to the misalignment of the crosshair with the wrong graduation. This error adds 0.600 m to all subsequent...
59